Miko

Miko is an AI companion for Windows that pairs local conversation models with an animated 3D character on your desktop. It's for people who want spoken conversations with a customizable avatar. The Unity avatar moves its lips with speech, blinks and performs gestures in a transparent overlay that lets clicks pass through to other applications. You can replace the character with your own VRM model.

Ollama runs the language model locally, with Qwen 2.5, Llama 3 and dolphin-llama3 among the named choices. You can speak through push-to-talk or continuous microphone listening, or type messages into a console. Speech recognition uses RealtimeSTT and Whisper. Replies can start playing before the model finishes writing the full response.

The voice engine determines where speech generation happens. GPT-SoVITS provides offline voice cloning on your hardware, keeping that processing local; an NVIDIA GPU with CUDA is recommended. Fish Audio generates speech in the cloud and requires an internet connection and API key. It supports multilingual, expressive speech without using local GPU memory for voice synthesis. The language model still runs through Ollama when you choose cloud speech.

Miko saves conversation history locally and recalls relevant past exchanges. Its memory system gives preferences, names and other important facts more weight, reduces the prominence of older memories and suppresses repeated statements. It also supports Tuya smart-home device control. The application runs on 64-bit Windows 10 and 11, and its open-source code uses the GPL-3.0 license.